Using the python package on longer audio

[JRMeyer]

>>> saseptim
[March 1, 2018, 11:04am]

Hi

The documentation states that the python package can be used to run the
pretrained model on clips of around 5 seconds. Why is there such a
limitation? Librivox samples can run up to 20 seconds long.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/using-the-python-package-on-longer-audio]

[JRMeyer]

>>> kdavis
[March 2, 2018, 3:12pm]

The neural net architecture used and the GPU memory are the limiting
factors and are independent of the Librivox sample size.

[Archived Post]

[JRMeyer]

>>> saseptim
[March 4, 2018, 10:50am]

So basically, if I have a GPU or CPU with lots of memory, then I can run
on more than 5 seconds?

[Archived Post]

[JRMeyer]

>>> kdavis
[March 4, 2018, 10:58am]

No. As mentioned, the neural net architecture used is also a limiting
factor. BRNN can not handle arbitrarily long sequences well.

You can experiment with longer audio clips but the recognition quality
will likely suffer.

[Archived Post]

[JRMeyer]

>>> saseptim
[March 4, 2018, 11:04am]

I see, but in order to achieve such good results on Librivox, you must
be able to get good results on its samples which are up to 35 seconds
long, so how does that work? slash
Also, I noticed that zero-padding the input audio signal hurts the
results, is this characteristic of the BRNN?

[Archived Post]

[JRMeyer]

>>> kdavis
[March 4, 2018, 12:06pm]

We get good result on the Librivox clean test set.

Some of the Librivox clean test samples are longer than 5secs; some
shorter. 5 sec is not a hard and fast rule. However, longer clips will
run in to the problem I mentioned, BRNN's can not handle arbitrarily
long sequences well.

[Archived Post]